Class BPNet

Hierarchy

BPNet

Index

Constructors

constructor

Properties

Methods

Constructors

constructor

new BPNet(shape: NetShape, opt?: BPNetOptions): BPNet

- Defined in bpnet.ts:25
Parameters
- shape: NetShape
- opt: BPNetOptions = {}
Returns BPNet

Properties

b

偏值

hlayer

hlayer: number

隐藏层层数

mode

mode: Mode = 'sgd'

梯度更新模式

rate

rate: number = 0.01

学习率

Optional scale

scale: Matrix

缩放比

Readonly shape

shape: NetShape

w

权值

Methods

adjust

adjust(dy: Matrix[], dw: Matrix[]): void

- Defined in bpnet.ts:197
调整权值偏值
- w = w - α * (∂J / ∂w)
- b = b - α * (∂J / ∂hy)
Parameters
- dy: Matrix[]
- dw: Matrix[]
Returns void

af

af(l: number): undefined | ActivationFunction

- Defined in bpnet.ts:59
获取当前层激活函数

Parameters
- l: number
Returns undefined | ActivationFunction

backPropagation

backPropagation(hy: Matrix[], xs: Matrix, ys: Matrix): { dw: Matrix[]; dy: Matrix[] }

- Defined in bpnet.ts:146
多样本求导，计算单个样本倒数求和，然后计算平均倒数

Parameters
- hy: Matrix[]
- xs: Matrix
- ys: Matrix
Returns { dw: Matrix[]; dy: Matrix[] }
- dw: Matrix[]
- dy: Matrix[]

bgd

bgd(xs: Matrix, ys: Matrix, opt: TrainingOptions): Promise<void>

- Defined in bpnet.ts:227
标准梯度下降，全部样本导数的平均值

Parameters
- xs: Matrix
- ys: Matrix
- opt: TrainingOptions
Returns Promise<void>

calcBackPropagation

calcBackPropagation(hy: Matrix[], xs: Matrix, ys: Matrix): { dw: Matrix[]; dy: Matrix[] }

- Defined in bpnet.ts:177
单样本求导，对每个输出单元的求导，对每个权重的求导
- J = 1 / 2 * (hy - y)^2
- ∂J / ∂hy[last] = (1 / 2) * 2 * (hy - y) = (hy - y) * hy[激活函数求导]
- ∂J / ∂w = dy.T * lastHy
- ∂J / ∂hy[now] = (nextDy * nextW) * hy[激活函数求导]
链式求导法则：
- 反向的计算过程；
- 分支节点相加；链路节点相乘法；
- 如有激活函数需乘激活函数的导数；
Parameters
- hy: Matrix[]
- xs: Matrix
- ys: Matrix
Returns { dw: Matrix[]; dy: Matrix[] }
- dw: Matrix[]
- dy: Matrix[]

calcLoss

calcLoss(xs: Matrix, ys: Matrix): number

- Defined in bpnet.ts:217
计算一组样本的当前损失

Parameters
- xs: Matrix
- ys: Matrix
Returns number

checkInput

checkInput(xs: Matrix): void

- Defined in bpnet.ts:305
Parameters
- xs: Matrix
Returns void

checkOutput

checkOutput(ys: Matrix): void

- Defined in bpnet.ts:311
Parameters
- ys: Matrix
Returns void

checkSample

checkSample(xs: Matrix, ys: Matrix): void

- Defined in bpnet.ts:317
Parameters
- xs: Matrix
- ys: Matrix
Returns void

cost

cost(hy: Matrix, ys: Matrix): number

- Defined in bpnet.ts:207
平方差带价函数多输出求平均值
- J = 1 / 2 * m * ∑m(hy - ys) ** 2
Parameters
- hy: Matrix
- ys: Matrix
Returns number

fit

fit(xs: Matrix, ys: Matrix, opt?: Partial<TrainingOptions>): Promise<void>

- Defined in bpnet.ts:325
Parameters
- xs: Matrix
- ys: Matrix
- opt: Partial<TrainingOptions> = {}
Returns Promise<void>

forwardPropagation

forwardPropagation(xs: Matrix): Matrix[]

- Defined in bpnet.ts:100
计算整个网络输出
- layer[hy][t-1] * w[t] + b
- hy = θ1 * X1 + θ2 * X2 + ... + θn * Xn + b
Parameters
- xs: Matrix
  
  inputs
Returns Matrix[]

mbgd

mbgd(xs: Matrix, ys: Matrix, opt: TrainingOptions): Promise<void>

- Defined in bpnet.ts:272
批量梯度下降，多个样本导数的平均值

Parameters
- xs: Matrix
- ys: Matrix
- opt: TrainingOptions
Returns Promise<void>

predict

predict(xs: Matrix): Matrix

- Defined in bpnet.ts:127
预测最后一层输出

Parameters
- xs: Matrix
Returns Matrix

predictNet

predictNet(xs: Matrix): Matrix[]

- Defined in bpnet.ts:135
预测整个网络，包含输入层

Parameters
- xs: Matrix
Returns Matrix[]

scaled

scaled(xs: Matrix): Matrix

- Defined in bpnet.ts:114
按照缩放比缩放新的特征

Parameters
- xs: Matrix
Returns Matrix

sgd

sgd(xs: Matrix, ys: Matrix, opt: TrainingOptions): Promise<void>

- Defined in bpnet.ts:246
随机梯度下降，单个样本的导数

Parameters
- xs: Matrix
- ys: Matrix
- opt: TrainingOptions
Returns Promise<void>

toJSON

toJSON(): string

- Defined in bpnet.ts:67
Returns string

json字符串

unit

unit(l: number): number

- Defined in bpnet.ts:51
获取当前层单元数

Parameters
- l: number
Returns number

Static fromJSON

fromJSON(json: string): BPNet

- Defined in bpnet.ts:82
加载json化的模型

Parameters
- json: string
Returns BPNet

BPNet

Generated using TypeDoc